Hybrid Parallel Twin-Tower Model Advances Arrhythmia Detection in ECG Signals: Integrating Convolutional Neural Networks and Kolmogorov–Arnold Networks

Background: Electrocardiogram (ECG) signal classification plays a crucial role in the early diagnosis and prevention of arrhythmias. However, the diversity of ECG signals, data imbalance, and the limitations of traditional models in generalization pose significant challenges to classification accuracy. Therefore, an effective algorithm is urgently needed to address these issues. Objective: This study presents an innovative hybrid parallel twin-tower model (CNN–KAN) that integrates the advantages of convolutional and interpretable neural networks to enhance the accuracy of ECG signal classification. Method: The CNN–KAN introduced adopts a parallel twin-tower architecture. The CNN automatically extracts the temporal features of ECG signals, while the KAN focuses on capturing deep structural features. The features of the CNN and KAN are cleverly fused in parallel to sense ECG signals and identify heart rhythm patterns. To enhance classification accuracy, strategies such as balancing data samples and sorting features based on variance are utilized. In addition, systematic experiments demonstrate the advancement, stability, and generalization ability of CNN–KAN. Results: On the MIT–BIH dataset, the average classification accuracy of CNN–KAN was 97.77%. The average classification accuracy under sample balancing was 99.25%. After further incorporating variance feature sorting, the optimized classification accuracy increased to 99.79%. The above results indicate that CNN–KAN achieves strong ECG classification performance under the heartbeat-level partitioning scheme predefined by Kaggle. Additionally, PTB experiments further support the applicability of this method to another ECG dataset and classification task. Conclusions: This study demonstrates that the CNN–KAN integrates multi-view and multi-scale features through a parallel twin-tower architecture. This architecture provides a method for automated ECG heartbeat classification; however, its inter-patient generalization capability and clinical applicability require further validation.